salyamq/kazspam-v1

KazSpam v1

KazSpam v1 is a Kazakh-focused dataset for spam and unwanted-content classification.

The dataset contains web-crawled texts annotated into four categories:

• normal — ordinary useful or neutral text;
• casino — gambling, casino, betting, bookmaker, and related promotional content;
• porn — pornography, adult content, and sexual-service advertisements;
• other — other spam, advertising, promotional, or unwanted content.

github repo: https://github.com/salyamq/kaz-datasets

Dataset Summary

Property	Value
Number of examples	10,000
Language focus	Kazakh
Labels	4
Data format	JSONL
Annotation type	Automatic
Release version	v1

Label Distribution

Label	Examples
casino	7,071
normal	1,460
porn	1,424
other	45

Features

Each record contains:

• text — input document;
• classification — assigned category;
• source — original dataset source;
• category — original document length bucket;
• model — model used for automatic annotation.

Example:

{
  "text": "Мәтін мысалы...",
  "classification": "casino",
  "source": "culturax",
  "category": "700-1500_tokens",
  "model": "inclusionai/ling-3.0-flash"
}

Data Sources

The sample was collected from sources: stukenov/sozkz-corpus-clean-v3 (apache 2.0 license), which contains

• culturax
• madlad400
• hplt_new
• mc4
• cc100
• kazparc_sync
• md_leipzig
• md_kazakhNews
• md_oscar
• moscar
• kazparc
• kazsandra
• md_kazakhBooks
• wikipedia
• belebele
• sib200
• wikiann

Annotation

The labels were generated automatically using a language model with the following classification scheme:

1. normal
2. casino
3. porn
4. other

The dataset should be treated as automatically annotated. Some examples may contain ambiguous or noisy labels, especially when a document combines ordinary text with spam fragments.

Intended Uses

This dataset can be used for:

• training spam classifiers;
• detecting gambling and adult-content spam;
• content moderation experiments;
• evaluating multilingual and Kazakh-language classifiers;
• filtering web-crawled text corpora.

Limitations

• The dataset is not a perfectly balanced classification benchmark.
• Most examples belong to the casino class.
• Texts may contain Russian, English, machine-translated, or corrupted fragments.
• The normal label does not guarantee that the text is high-quality or free from harmful content.

Citation

@misc{salyamq_kazspam_v1,
  author       = {{salyamq}},
  title        = {KazSpam v1},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/salyamq/kazspam-v1}}
}